Skip to main content
HomeBlogHow-To Guides
Back to all articles
How-To Guides

How to Write Scripts for AI Voice Agents That Actually Convert

Rahul AgarwalOctober 19, 202611 min read
ai voice agent scriptsvoice agent prompt writingai call scriptconversational ai scripting

How to Write Scripts for AI Voice Agents That Actually Convert

The difference between an AI voice agent that feels helpful and one that frustrates callers comes down almost entirely to how it's scripted. The underlying technology — Deepgram for speech recognition, ElevenLabs for voice, Claude or GPT-4 for language — is broadly consistent across quality platforms. What varies dramatically is the quality of the persona, knowledge base, conversation flows, and escalation logic that humans write on top of it.

This guide covers everything you need to write AI voice agent scripts that sound natural, handle edge cases gracefully, and achieve your business objectives.


Part 1: The Mental Model — Writing for Ears, Not Eyes

The single biggest mistake people make when scripting AI voice agents is writing text as if it will be read. Voice and text are fundamentally different:

Text is scanned. Voice is heard. Readers jump between paragraphs and headings. Callers process audio linearly, in real time, with no ability to re-read.

Text can be dense. Voice must be simple. A 3-sentence paragraph that reads cleanly becomes incomprehensible when spoken at conversational speed.

Text has no pacing. Voice has rhythm. Good voice scripts include natural pause points, questions that invite response, and confirmation loops that ensure the caller is still engaged.

The 5-Second Rule

If any single AI response takes more than 5 seconds to speak aloud, it's too long. Read your scripts out loud, stopwatch running. Anything over 5 seconds needs to be split, shortened, or restructured as a question that the caller answers before the AI continues.

Bad (too long):

"Thank you for calling Apex Home Services, where we've been serving the greater Phoenix metro area for over 20 years with HVAC, plumbing, and electrical services. My name is Emma, and I'm here to help you with whatever you need today. Could you tell me a little bit about what you're calling in for today?"

Good:

"Thanks for calling Apex Home Services. I'm Emma, your AI assistant. What can I help you with today?"

The second version is friendlier, easier to process, and invites a response immediately. The caller doesn't need the 20-year history in the greeting.


Part 2: Persona Setup — The Foundation of Everything

Before writing a single script line, define your agent's persona. This is the single document that governs tone, language, and behavior across every interaction.

The 7 Elements of an Effective AI Persona

1. Name Give your agent a real first name. Not "APEX AI" or "Virtual Assistant." A name creates rapport immediately. Common recommendation: choose a name that's common enough to feel neutral but not so common that it's cliché. Emma, Alex, Jordan, Maya — these work across most industries.

2. Tone Pick two to four adjectives that describe how the agent should sound. Examples:

  • Medical practice: Calm, reassuring, professional, efficient
  • Car dealership: Friendly, knowledgeable, low-pressure, enthusiastic
  • Collections: Respectful, factual, patient, solution-focused
  • Real estate: Energetic, helpful, knowledgeable, responsive

3. Speaking Pace Should the agent speak quickly or more slowly? For complex information (insurance, medical), slower is safer. For appointment booking, faster keeps things efficient.

4. What It Will and Won't Discuss Define the scope explicitly. "Emma handles appointment scheduling, service information, after-hours emergencies, and appointment reminders. Emma does not provide pricing beyond published rates, does not make warranty decisions, and does not handle billing disputes."

5. Escalation Trigger Under what conditions should the AI offer to transfer? Define them clearly:

  • Caller requests human three times
  • Caller expresses anger or frustration (detected by LLM)
  • Caller asks a question outside defined scope
  • Emergency criteria met (define what "emergency" means for your business)

6. Disclosure Language How does the agent identify itself as AI when asked? This is legally required in many jurisdictions and good practice everywhere. Example: "I'm Emma, an AI assistant. I handle most questions quickly, but I can connect you with a team member if you need something specific."

7. Closing Behavior How does the agent end calls? "Is there anything else I can help you with today? ... Perfect. You'll receive a confirmation text shortly. Thank you for calling Apex, and have a great day."


Part 3: Knowledge Base Structure

The knowledge base is the AI's source of truth — the content it draws on when answering questions. Structure matters enormously.

What to Include

Tier 1: Universal (every agent needs this)

  • Company name, address, hours, phone number
  • Primary services or products (plain language descriptions)
  • Pricing — either specific figures or "how to find pricing"
  • Service area / eligibility criteria
  • Standard escalation options (who to transfer to, how)

Tier 2: Call-Type Specific For each major call type your agent handles, provide:

  • How to recognize this type of call (intent signals)
  • What information to collect
  • What to do with that information
  • How to confirm and close

Example for scheduling:

"When a caller wants to book a service appointment, collect: (1) name, (2) service address, (3) type of service needed (HVAC, plumbing, or electrical), (4) preferred date/time window, (5) contact phone/email for confirmation. After collecting, check available slots in ServiceTitan and offer three options. Confirm selected option, send SMS confirmation, and close with: 'You'll receive a confirmation text at [number]. Is there anything else I can help you with?'"

Tier 3: FAQ Bank A structured list of common questions and preferred answers. Format each as:

Q: [Exact or paraphrased question as caller would ask it]
A: [Exactly how the agent should respond — in spoken voice, not written text]

Target 30–50 Q&A pairs for a well-configured agent. Below 15, you'll have too many fallbacks. Above 100 without good structure, response quality degrades.

The "Preferred Answer" Principle

Don't just tell the AI what's true — tell it exactly how you want it said. The difference between "Our warranty is 1 year on parts and labor" and the preferred answer matters enormously:

Preferred answer: "All of our parts and labor come with a one-year warranty. If anything goes wrong within that year, just call us and we'll come back out at no charge. That's our promise."

The second version is warmer, more specific, and ends with a confidence signal. Write your FAQ answers as your best human agent would actually speak them.


Part 4: Conversation Flow Design

A conversation flow is the map of how a call moves from opening to close, including branching logic for different caller types and situations.

The Core Call Structure

Every call follows the same basic structure:

Opening → Intent Detection → Information Gathering → Action/Resolution → Confirmation → Close

Design each stage intentionally:

Opening (5 seconds) Goal: Establish identity, create immediate comfort, invite the caller to speak. Format: "[Greeting] + [Company name] + [Agent name] + [Open question]" Example: "Hi, thanks for calling Apex Home Services. I'm Emma. What can I help you with today?"

Intent Detection This is handled by the LLM automatically, but you can improve it by:

  1. Listing the exact phrases that indicate each intent in your knowledge base
  2. Providing examples of ambiguous phrasings (e.g., "I need to change something" might mean reschedule, update address, or change service type)
  3. Instructing the agent to clarify ambiguous intent before proceeding: "I want to make sure I help you with the right thing — are you looking to schedule a new appointment or update an existing one?"

Information Gathering Write gathering sequences as a series of single, clear questions — never multiple questions in one turn.

Bad:

"Can I get your name, address, and what type of service you need, and what date works for you?"

Good (sequential):

"Can I get your name?" [Caller responds] "And what's the service address?" [Caller responds] "What type of service are you looking for — HVAC, plumbing, or electrical?"

Action/Resolution After gathering, the AI takes action and narrates it: "Let me check available times for Tuesday the 15th... I have openings at 9am, 1pm, and 3pm. Which works best for you?"

Confirmation Always read back key information before closing: "Just to confirm — I've scheduled a plumbing service visit for John Martinez at 42 Oak Street on Tuesday the 15th at 1pm. You'll get a confirmation text at 602-555-1234."

Close Consistent, warm, brief: "Is there anything else I can help with? ... Perfect. Thank you for calling Apex. Have a great day."


Part 5: Objection and Edge Case Handling

Most scripts fail in edge cases — the moments when callers say something unexpected. Here's how to handle the most common ones.

"I don't want to talk to a robot"

Never argue with this. Never say "I'm actually very helpful." Acknowledge and offer the alternative immediately.

"Of course — I completely understand. Let me connect you with one of our team members. Just a moment."

If a human isn't available: "Our team is currently [unavailable / helping other customers]. I can take a message and have someone call you back within [timeframe], or you can call back at [hours]. Which would you prefer?"

"How much does it cost?"

This depends on whether you've published pricing. Two approaches:

If you have published pricing:

"Our [service] typically runs between $[low] and $[high] depending on the scope. The technician will give you an exact quote when they're on-site. Would you like to schedule a visit?"

If pricing requires an assessment:

"Pricing for [service] depends on a few factors we'd need to assess in person. There's no charge for the estimate. Would you like to schedule a free assessment?"

Never let pricing questions become dead ends. Always bridge back to the next step.

"I need to speak to a manager"

Don't escalate to "manager" unless a manager is available and appropriate. Instead:

"Absolutely. Our customer experience team handles concerns directly. I can connect you right now if they're available, or schedule a callback at a time that works for you. Which would you prefer?"

"This is an emergency"

Define "emergency" in your knowledge base with specific examples. For home services: active water damage, gas smell, no heat in freezing temperatures, electrical panel sparking.

"That sounds like an emergency — let me make sure we get someone to you quickly. Can you confirm your address? [confirms] I'm alerting our on-call technician right now. You'll receive a call or text within 15 minutes. Please stay on the line if the situation is immediately dangerous."

Never handle an emergency with "I'll have someone call you back."

"I already called about this"

"I'm sorry there's been a delay. Let me pull up your information. Can I get your name or the address associated with the service request?"

After pulling details: "I see [situation]. Let me connect you with our team to get this resolved quickly."


Part 6: Writing for Specific Call Types

Outbound Appointment Reminders

The goal is confirmation, not conversation. Keep it short.

"Hi, this is Emma from Apex Home Services with a reminder about your plumbing service appointment on [day] at [time]. To confirm, press 1 or say 'confirm.' To reschedule, press 2 or say 'reschedule.' To cancel, press 3."

After confirmation:

"Great, you're confirmed for [day] at [time]. Your technician will call 30 minutes before arrival. Is there anything else I can note for them before the visit?"

Outbound Lead Follow-Up

The tone shifts to consultative, not transactional.

"Hi [name], this is Emma from Apex Home Services. You submitted a request about [service] on [date]. I wanted to make sure someone reached out to you. Do you still have questions about [service], or are you ready to schedule?"

Keep it warm, one question at a time. The goal is to get a "yes" to scheduling, not to pitch features.

New Customer Intake

Collect more, talk less. Your job is to get information efficiently.

"Hi, I'm Emma from Apex Home Services. Looks like this is your first time calling us — welcome! I'll get you set up quickly. Can I start with your name?"

Proceed through intake fields methodically, confirming each one before moving to the next.


Part 7: Testing Your Script Before Going Live

Never launch an AI voice agent without testing these 10 scenarios:

ScenarioWhat to Test
1. Standard bookingDoes the agent collect all required fields and book correctly?
2. Reschedule requestDoes it find the existing appointment and offer alternatives?
3. "How much does it cost?"Does it give the right pricing response without dead-ending?
4. Emergency callDoes it escalate correctly and immediately?
5. "I want a human"Does it offer transfer or callback without arguing?
6. Caller speaks quietly/mumblesDoes it ask for clarification gracefully?
7. Long silenceDoes it prompt the caller correctly?
8. Caller goes off-topicDoes it acknowledge and redirect without frustration?
9. Multiple requests in one turnDoes it address each request in order?
10. "Are you a real person?"Does it answer honestly and immediately?

For each scenario, evaluate:

  • Did the AI understand the intent correctly?
  • Was the response the one you would have written for a best human agent?
  • Did it feel natural when heard aloud?
  • Did it achieve the goal of the interaction?

Common Scripting Mistakes to Avoid

1. Over-explaining. Every extra word is a second the caller isn't speaking. Ruthlessly cut everything that doesn't serve the immediate goal.

2. Corporate language. "We appreciate your patience" → "Thanks for waiting." "Your call is important to us" → Don't say this at all. "Please be advised" → [Delete entirely]

3. Passive escalation. "You can also speak with a human if you prefer" (weak) → "I can connect you right now — would that help?" (active, caring)

4. Ending questions that invite "no." "Is there anything else I can help you with?" invites "no" and ends the call. That's fine for closing — but don't use this phrasing in the middle of an interaction.

5. No handling for the unexpected. Every script needs a catch-all for questions outside scope: "That's a good question — I want to make sure you get the right answer on that. Let me connect you with our team or have someone call you back. Which would you prefer?"


Final Tip: Iterate From Real Calls

Your first script won't be your best. After going live, review 10–20 call transcripts per week for the first month. Look for:

  • Moments where the AI gave the wrong answer
  • Questions that appear frequently but aren't in your FAQ
  • Calls that escalated that shouldn't have
  • Calls that didn't escalate that should have

Update your knowledge base based on what you find. Most well-maintained AI agents improve materially within 60 days of launch.


Ready to put these scripting principles to work? Start a QuickVoice free trial — the setup wizard walks you through persona, knowledge base, and flow configuration step by step, with templates for 12 industries.

R
Rahul Agarwal
Writing about AI voice, business automation, and the future of customer communication at QuickVoice.

Ready to deploy AI voice for your business?

No code. No credit card. First agent live in under 30 minutes.